Safety evals need stronger methods before model scores guide deployment
Anthropic says safety evaluations should be rigorous enough to decide whether a model is ready to deploy, which safeguards should ship, and what safety training comes next. The evaluations test whether a model’s response makes harm more or less likely, but results can be gamed, measure the wrong thing, or fail to reproduce. Drawing on reforms such as preregistration and better measurement, the post recommends a more disciplined approach, starting with teams that combine technical and domain expertise.